1F-05 - Development and validation of gestational age estimation algorithms for non-live births in administrative healthcare databases

Background: Non-live births—spontaneous abortions, induced abortions, and stillbirths—are a crucial consideration in pregnancy studies utilizing administrative healthcare databases but are frequently overlooked due to the lack of reliable data on pregnancy start dates.

Objectives: To develop and validate claims-based algorithms for estimating gestational age at the time of non-live births.

Methods: Using the National Health Insurance Database of South Korea, we established a reference standard by linking gestational week data from influenza vaccination records in the Korea Disease Control and Prevention Agency registry (Jan 2018–Jun 2022). A hierarchical algorithm was developed to identify pregnancy episodes, and non-live births were stratified into spontaneous/induced abortion and stillbirth groups. Four approaches were tested to estimate the gestational age: (1) assigning pregnancy outcome-specific gestational ages, (2) adjusting gestational age based on 29 gestational markers, (3) fitting a regression model with a least absolute shrinkage and selection operator (LASSO) approach for predictor selection, and (4) using a random forest model. Algorithms were evaluated by the proportion of estimates falling within 1–4 weeks of the reference standard and the mean squared error (MSE). External validation was conducted using an independent dataset of pregnancies achieved through medically assisted reproduction.

Results: After applying the hierarchical algorithm, a total of 4102 cases of spontaneous/induced abortions and 379 cases of stillbirths were identified for model development. The random forest model performed best for predicting gestational age for both spontaneous/induced abortions (MSE: 1.60 weeks²) and stillbirths (MSE: 3.05 weeks²), with 91.2% (95% CI 90.3%–92.0%) and 88.4% (95% CI 84.8%–91.2%) of predictions falling within two weeks of the reference standard, respectively. In the external validation set, the gestational marker-based adjustment approach was the most accurate for spontaneous/induced abortions (MSE: 9.67 weeks²; % within 2 weeks: 75.2%, 95% CI 74.2%–76.1%), while the random forest model performed best for stillbirths (MSE: 13.43 weeks²; % within 2 weeks: 69.0%, 95% CI 61.7%–75.4%).

Conclusions: These algorithms provide reliable methods for estimating the gestational age of non-live births, supporting pregnancy research within the Korean population. By addressing selection bias and exposure misclassification when including non-live births in pregnancy studies, these tools can enhance the reliability of studies on medication safety and maternal-fetal outcomes.